Your best process lives in one person's head
Every company has a handful of tasks with a right way to do them. The quote that never leaves money on the table. The onboarding email sequence that actually gets clients to show up prepared. The weekly report the CEO reads instead of skims. And in most businesses, the "right way" lives in exactly one place: the head of whoever does it best.
AI has quietly changed what that knowledge is worth. Most serious AI tools now support some version of a reusable playbook — a written procedure the AI pulls up whenever a task calls for it, and then follows step by step. The names vary by platform (skills, instructions, recipes), but the concept is portable: stop hoping the AI does it your way, and hand it the way.
A playbook is not a briefing
We've written about standing context — the permanent brief that tells AI who you are, who your customers are, how you sound. A playbook is the other half:
- The briefing covers who we are. It applies to everything.
- A playbook covers how we do one specific job. It applies when that job comes up — and it carries the steps, the rules, and the quality bar.
"Write it in our voice" comes from the briefing. "A quote always includes three tiers, names a specific start date, and never discounts past 15% without flagging it" — that's the quoting playbook. You need both, and they do different work.
The temp test
Here's the standard a playbook has to meet, for AI and humans alike: could a competent temp follow it on day one, with nobody standing behind them? If yes, an AI can follow it too. If no — if the steps secretly require knowing "what we usually do" — the playbook isn't done, no matter how official it looks.
A playbook that passes usually has five parts:
- When to use it. "Use this whenever a prospect asks for pricing" — so the right procedure fires at the right moment.
- The steps, in order. What to gather first, what to produce, what to check before it's done.
- The hard rules. The non-negotiables, stated as non-negotiables: floors, approvals, things that always get a human look.
- One gold example. A real, finished instance of the task done right. Nothing communicates the standard faster.
- What "done" looks like. The checklist the output must pass before anyone sees it.
The fastest way to write one: don't describe the process from memory. Take the last time the task went well, and narrate what actually happened, step by step. Real instances beat idealized flowcharts every time.
What you get for the effort
- Consistency without supervision. The tenth quote looks like the first. Output quality stops depending on who asked the AI, or how carefully they prompted.
- Onboarding collapses. A new hire with your playbooks — and an AI that follows them — does credible work in week one, because "how we do things" is no longer three months of osmosis.
- The bus factor drops. When your best closer's quoting instincts are written down and executable, they stop being a single point of failure.
And notice the loop from our iteration post: when output misses, you now fix the playbook, not the draft. Every correction compounds instead of evaporating when the chat ends.
The quiet second payoff
A playbook an AI can follow is, almost word for word, the specification for automating that task outright. When we build automated workflows for clients, the expensive early phase is extracting exactly this — the real steps, the real rules, the real quality bar — out of people's heads. Companies that have already written their playbooks skip most of that. Companies that haven't, pay for it during the build.
So the effort pays twice: better, more consistent AI output today — and a shovel-ready spec for the day that task should run on a trigger instead of a request.
This week
- Pick the task you've explained to a human most often. That frequency is the signal.
- Write the one-page playbook: when it applies, the steps, the hard rules, one gold example, the done-checklist.
- Give it to your AI with a real instance of the task, and grade the output against your example.
- When something misses, fix the playbook and run it again. Three rounds usually gets it dependable.
Then hand the same playbook to a teammate and watch it work for them too. That's the tell you've written something durable.
If you've got playbooks written and you're wondering which one deserves to become a fully automated workflow, that's a conversation we're good at. Book a 30-minute call and bring your best one.